np.timedelta64 integer coercion in older numpy

[jbrockmendel]

Ref #19333

In sufficiently old numpy (the build that caught this bug was running 1.9.2) timedelta64 comparisons have unexpected behavior:

value = np.timedelta64(1, 'ns')
new_value = int(value)
assert value == new_value   # <-- raises with modern numpy, passes with older versions

This particular comparison is done in the RangeIndex constructor, the result being that sometimesRangeIndex(value, value*10, value) returns RangeIndex(1, 10, 1) instead of raising.

The immediate problem can be fixed by checking more carefully in the RangeIndex constructor. No idea if there are other places that are affected.

---

A related behavior that I find unhelpful that is present with 1.14.1 is:

>>> td = pd.Timedelta(days=1).to_timedelta64()
>>> pd._libs.lib.is_integer(td)
True

[jreback]

what is the purpose of opening a separate issue for this? this is a well known detail of how np.timedelta64 are implemented. which is True for all numpies (the 2nd part)

[jbrockmendel]

The purpose is to address the behavior. The #19333 addresses one effect of this behavior, but does not fix it, doesn't even fix RangeIndex.__new__

As for the second part, lib.is_integer is part of pandas. If we want it to behave in a more reasonable way, that's easy to make happen.